ABSTRACT
Chapter 8 presents the front-door theorem and its use to adjust for confounding via the front-door method, which targets the population averaged treatment effect. The chapter also extends the method to estimate the average effect of treatment on the treated. Assumptions required for validity of the method are framed in terms of causal directed acyclic graphs and also in terms of potential outcomes. Proof of validity of the method is given in terms of potential outcomes. Surrogate markers, including a perfect surrogate and a partial surrogate, are introduced to motivate the front-door method. It is argued that in much of science, we have more information on the total effect of a cause on an outcome than we have on its mechanism, which could account, in large part, for the currently limited application of the front-door method. However, examples are presented that suggest it may find use in post-marketing follow-up of randomized clinical trials relying on surrogate marker outcomes or in post-marketing safety analyses. R code is also provided.
